Why general-purpose AI falls short for legal drafting
Why general-purpose AI falls short for legal drafting is not primarily a question of intelligence. It is a question of context, and context is exactly what broad AI assistants lack.
Trained on everything, specialised in nothing
A general-purpose AI assistant has been trained on a vast corpus of public text: news articles, forum posts, academic papers, contracts from different jurisdictions and different industries, template libraries built for markets your clients will never encounter. The result is a system that can produce plausible-sounding legal language without reliably producing correct legal language for your specific practice.
The gap matters more than it sounds. A share purchase agreement for a UK mid-market transaction has particular conventions around conditions, completion mechanics, locked-box versus completion accounts and warranty packages. A broad model that has been trained on a statistical average of every contract it has ever seen will drift toward that average rather than reflecting what is actually standard in your market and for your clients.
It knows nothing specific about your firm
The more fundamental problem is that a general-purpose assistant has no information about your firm. It does not know your standard position on limitation of liability, the clause variations your most important client expects, or the counterparty you have negotiated with several times whose opening position you know to be flexible by a significant margin.
Every session starts fresh. You can feed context through a prompt, but that context disappears when the session ends. There is no accumulation of knowledge, no firm-owned record of positions and outcomes. You are paying for a tool that gets no smarter about your practice over time. That is not a tool that compounds in value; it is a tool that plateaus.
Hallucination risk is higher without firm-specific grounding
General-purpose models produce errors in legal contexts, and the errors tend to be confident ones. A model without access to your actual precedents will occasionally invent clause language, references or defined terms that look reasonable at a glance but are not accurate. When output is grounded in your real documents, the scope for that kind of error is meaningfully reduced, because the model is drawing on what you have actually used and approved rather than what it has statistically inferred.
What a purpose-built approach does differently
Scriptum drafts from your firm's precedents, works inside Microsoft Word so the context of the document in progress is always available, and captures a structured record of your firm's positions over time. The result is a drafting assistant that becomes more useful the longer your team uses it, rather than one that resets at the start of every new session.
To see the difference in practice, book a demo and we can demonstrate with a document type your team handles regularly.